Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++

[Engine] SYCL: a state dump for the pipelined windows, and the two defects it already shows

Niko1221478964ecaa78
diff --git a/sycl/src/program/generate.cpp b/sycl/src/program/generate.cpp
index 20d912779..10da48053 100644
--- a/sycl/src/program/generate.cpp
+++ b/sycl/src/program/generate.cpp
@@ -11540,6 +11540,31 @@ int main(int argc, char **argv) try {
B.launched ? "launched" : B.ready ? "ready" : "none", on ? " ON" : "");
x = outp[(size_t) a];
p = A.p + a + 1;
+ // STRATA_PIPELINE_DUMP=<file>: after every window, append a hash and the sum of the first stage's
+ // sequence state (its GDN rows: the recurrence state plus the conv history). Two arms that ran
+ // the same tokens must agree here; the first position where they do not is the state that
+ // escapes. Debug only: it waits for the stage's stream, so it distorts the timing.
+ static const char* pl_dump = std::getenv("STRATA_PIPELINE_DUMP"); // read once
+ if (pl_dump != nullptr) {
+ const int64_t gdn_floats = g.ssm_state_size * g.ssm_v_heads * g.ssm_state_size +
+ g.ssm_conv_channels * (g.ssm_d_conv - 1);
+ const size_t n = (size_t) ss.gdn_alloc * (size_t) gdn_floats;
+ std::vector<float> h(n);
+ (void) DPCT_CHECK_ERROR(s0->memcpy(h.data(), ss.gdn_state, n * sizeof(float)).wait());
+ uint64_t hs = 1469598103934665603ull;
+ double sum = 0.0;
+ for (size_t i = 0; i < n; ++i) {
+ uint64_t b = 0;
+ std::memcpy(&b, &h[i], sizeof(float));
+ hs = (hs ^ b) * 1099511628211ull;
+ sum += (double) h[i];
+ }
+ if (std::FILE* df = std::fopen(pl_dump, "a")) {
+ std::fprintf(df, "pos %lld windows %lld gdn %zu sum %.9g hash %016llx\n", (long long) p,
+ (long long) dec_windows, n, sum, (unsigned long long) hs);
+ std::fclose(df);
+ }
+ }
if (!last && !pl_adapt()) return die("an adaptive refill failed");
if (on && !last) {
++pl_on;